HAMBURG BIT-BOTS AND WF WOLVES TEAM DESCRIPTION FOR ROBOCUP 2019 - HUMANOID TEENSIZE

Hamburg Bit-Bots and WF Wolves
         Team Description for RoboCup 2019
              – Humanoid TeenSize –

    Marc Bestmann1 , Hendrik Brandt1 , Timon Engelke1 , Niklas Fiedler1 ,
Alexander Gabel2 , Jasper Güldenstein1 , Jonas Hagge1 , Judith Hartfill1 , Tom
Lorenz2 , Tanja Heuer2 , Martin Poppinga1 , Ivan David Riaño Salamanca2 , and
                                Daniel Speck1
                1
                    Department of Informatics, Universität Hamburg,
                    Vogt-Kölln-Straße 30, 22527 Hamburg, Germany
                                   info@bit-bots.de
                                    www.bit-bots.de
                      2
                         Ostfalia – University of Applied Sciences
                      On Exer 2d, 38302 Wolfenbüttel, Germany
                                  robo-wm@ostfalia.de
                               https://www.wf-wolves.de

      Abstract. This team description paper presents the developments made
      by the joint team Hamburg Bit-Bots & WF-Wolves. We present new soft-
      ware approaches we programmed and evaluated, like the Dynamic Stack
      Decider (DSD), our advances in image processing and our improvements
      to the walking engine. Additionally the newly developed foot pressure
      sensors as well as our progress towards a reliable and fast servo control
      are introduced. We, the joint team of Hamburg Bit-Bots and WF Wolves,
      hereby apply for participation in the RoboCup 2019 in Sydney, Australia
      in the Humanoid TeenSize Team Competition. The content of this
      paper is the same as the application for the KidSize [5].

      Keywords: RoboCup · Humanoid · Soccer · World Model

1   Introduction

The Hamburg Bit-Bots and WF Wolves are applying to participate as a joint
team in RoboCup 2019. In last year’s competition we benefited greatly from
our cooperation. We were able to exchange robots freely between leagues which
allowed us to have more robots in play at multiple games. We have expanded
our partnership further in terms of hard- and software. The Hamburg Bit-Bots
have moved on from the Minibot platform used in previous years to focus more
on improving the shared robot platform. The Hamburg Bit-Bots are a group of
Bachelor, Master and Ph.D. students, that are supported by the Department of
Informatics at the University of Hamburg. The WF Wolves are supported by
the University of Applied Sciences Ostfalia.

2       Hamburg Bit-Bots and WF Wolves

1.1   Previous Achievements
The Hamburg Bit-Bots are part of a marketing campaign celebrating the 100th
anniversary of the Universität Hamburg. Through the campaign, RoboCup itself
is heavily advertised in Hamburg, Germany.
    The Hamburg Bit-Bots developed the ImageTagger [11]. Currently multiple
teams (see [9][10]) use this tool. Our publicly available instance holds more than
280,000 public images from the SPL and Humanoid League. 270,000 Labels have
been manually created for these images.
    In recent years, the members of the Team published a number of papers [7],
[11], [12], [20], [21].

2     Research
2.1   Vision with Neural Networks
We developed and evaluated two different approaches to vision with neural net-
works. Both approaches are described respectively in [21] and [12]. Our efforts
spent for data collection and image labelling (described in [11]) helped us to
further improve both methods.

2.2   World Model
Currently we are working on separately filtering sensor data in a local and a
global layer.
     The filtering process in the global layer happens in the relative space of each
robot. A particle filter is applied onto detected obstacles and players. To increase
the amount of information fed into the filter when filtering ball measurements,
the whole heat map generated by the FCNN is transformed into relative space
and used as measurement in the filter.
     In the global filtering layer, the information generated by the whole team is
collected via the team communication module and filtered in a separate particle
filter. Thus, the self localization and detections of one robot can be corrected by
the others.

2.3   Behavior
We were using our own framework, the stack machine, for defining the robot
behavior since 2014. This year, we rewrote large parts of the stack machine and
renamed it to Dynamic Stack Decider. It consists of modules defining actions or
decisions which are written in Python. Each decision module has a finite set of
return values. Based on these, another decision or an action is pushed on the
stack as defined in a simple Domain Specific Language (DSL). This approach
allows fast changes in the game logic, without the need of touching the codebase.
Modules on the stack define the active modules and the current path in the
directed acyclic graph (DAG) defined in the DSL. The different modules have
access to a shared blackboard to communicate with each other.

Hamburg Bit-Bots and WF Wolves TDP 2019            3

2.4   Motion

For the stabilization of our full body motions we are currently researching in
two directions. Firstly, we are trying to compute, based on sensory input from
foot pressure and IMU, secondary balance goals for inverse kinematics (IK) [19].
It returns the necessary joint values to reach the requested cartesian positions
while staying stable.
    Secondly, we are trying to transfer the promising results in recent deep re-
inforcement learning [17] onto our robot. Learned motion may have a higher
performance as well as requiring less tuning.

2.5   Walking

Last year, we managed to achieve a good walking on the artificial grass (see
section 4.6). Still, it needs to be improved for a higher speed and stability.
Since the performance of the current walk engine depends highly on the chosen
parameters, we evaluated different parameters in simulation. This allowed us not
only to find the best combination automatically, rather than trying out per hand,
it also gave us insight about the dependencies between different parameters, thus
allowing us to fine tune them better. Similar to the motions (see section 2.4) we
are also trying to apply deep reinforcement learning to this problem.

2.6   Hardware

In respect to research on our hardware, we are currently mostly working on
increasing the frequency of the control cycle to allow faster reactions to distur-
bances. While our current servo controller board (see section 3.1) works fine,
the cycle rate is limited due to the low performance of the micro controller and
the limited baudrate. We are currently replacing this with a never version using
an FTDI chip to increase the baudrate and split up communication into three
separate communication busses, one for each leg and one for the arms. Our test
results and calculations suggest that an update rate of 1 kHz can be achieved.
We will release this project as open-source hardware. Furthermore, the new ver-
sion will include two IMUs to improve the filtering for the estimated orientation
of the robot.
    Team Rhoban has developed a low cost pressure sensor to measure the ground
reaction forces [14]. We improved on their open source design in terms of speed
and resolution, see Section 3.2.

3     Hardware

Both teams (Hamburg Bit-Bots and WF Wolves) are using the same platform,
called Wolfgang (see 1b), which is an advancement of the platform used last year.
The main changes are in the electronics and sensors. The platform has the typical
20 DOF structure using Dynamixel MX-64 and MX-106. Three computers are

4       Hamburg Bit-Bots and WF Wolves

installed on it: an Intel NUC, an Nvidia Jetson TX2 and an Odroid XU-4.
They provide us with enough processing power to run FCNNs for the vision and
an evolutionary IK to create stable motions. The computers are connected via
GigaBit ethernet and the communication is handled via ROS message passing,
allowing a complete abstraction from the distributed system when programming
software.

3.1   Servo Control
In the last year we switched to the second version of the Dynamixel protocol.
This allows a more stable and faster control of the servos. Furthermore, it is
now possible to get a good estimate of the applied torque as well as using torque
based control. We are currently using the DXL Board from Rhoban [3] with an
upgraded firmware written by us, which supports the new protocol. This allows
for an almost error free and faster control of the servos than the CM730 we
and many other teams used in the past. We are currently reaching more than
200 Hz when reading and writing all servos. The software for controlling the
servos is based on an improved version of the DynamixelSDK and follows the
ros control [8] standards. This enables us to switch between position, velocity and
force control. It also provides the same interface as our simulation environment
in Gazebo [15], simplifying the transfer from simulation to reality.

3.2   Pressure Sensors

Fig. 1a: The hardware implementation of
the ForceFoot. In each of the four cor-
ners of the foot, a load cell sensor is
mounted. The sensor data is collected and
fed into the Dynamxel bus system by an
STM32F103 with an external analog to        Fig. 1b: The Wolfgang robot platform
digital converter.                          used by both teams.

Hamburg Bit-Bots and WF Wolves TDP 2019              5

Inspired by team Rhoban’s ForceFoot [4] we have improved on their design in a
number of ways (see 1a). We have increased the possible update rate of the load
cells from 80 Hz to a theoretical 9.5 kHz per cell. Since the data transfer rate
through the bus is limited we use filtering on the load cells output to achieve
a more stable measurement at a speed of 1 kHz. In the next months we will
scientifically compare our measurement of the center of pressure to an industrial
6 axis force torque sensor and publish our results. Our improved version of the
ForceFoot will be released as open-source hardware in the near future.

4     Software

The WF Wolves started to use ROS [18] in 2016 and the Hamburg Bit-Bots in
2017. We created a common set of messages for the RoboCup Soccer domain [6].
This allows us to easily share and exchange parts of our software. Our code is
open and available online3 .

4.1    Vision

Our vision is currently able to reliably detect and identify multiple object types
while running at about 10 frames per second on an Nvidia Jetson TX2. These
include the moving ball, the field with its markings and goalposts as well as
other robots (including team affiliation).
    The observation of the environment and accurate detection of various objects
is made possible by multiple specialized modules that form our vision. Each
module is responsible for identifying or classifying a specific aspect of the picture
for further feature extraction, creating a structure that allows our modules to
function mostly independently and only use other modules end results.
    This property is especially prominent for the obstacle detection module which
not only localizes obstructions but also provides potential candidates of other
robots and goalposts for the candidate module or goalpost module respectively.
The obstacle detection module itself also utilizes other modules’ results in order
to perform its task, like the horizon detection which is used to model the edges
of the field. The resulting horizon line is not completely straight and contains
multiple dents which is used for the obstacle detection that outputs a series of
candidate. These are used by the candidate module to assign team affiliation of
the observed robots by identifying the color of its team markers.
    For ball detection we have two alternative approaches. One approach is the
FCNN proposed in [21]. The other approach we are evaluating has been pro-
posed in [12]. The performance of this second approach was further improved by
replacing AlexNet with MobileNetV2 which is optimized for mobile platforms,
while the training process is already sped-up by using a finetuning approach
instead of full training. This can be used to train the network on competition in
a very short amount of time (in the order of five minutes).
3
    https://github.com/Bit-Bots, https://humanoid.wf-wolves.de/explore/

6       Hamburg Bit-Bots and WF Wolves

Fig. 2: Debug output of the vision system. The detected horizon is marked by a red
line, field markings are represented as a set of red dots. Around detected obstacles,
boxes are drawn (black: unknown obstacle, red: red robot, blue: blue robot, white:
white obstacles e. g. goalposts). The best rated ball candidate and discarded candidates
are indicated by a green circle or red circles respectively.

4.2   Team Communication

As in previous years, we are using the ”mixed team communication protocol”
(mitecom) [1]. This enables us to have robots from our teams playing together
as well as playing in a drop-in games. We already proved this ability in the last
years and hope that more teams will be using the standard in this years drop-in
challenge.

4.3   Localization

The main source of information for self-localizing on the field are the field
lines. The vision provides points on the image which belong to lines. Those
get transformed to positions relative to the robot and are put into the format
of a laser scan to be handled as if they were walls. This allows us to use the
standard particle filter based localization node provided by ROS (AMCL [13]).
It enables the robot to do global localization for example after pickup, when
no initial position is available (kidnapped robot problem). Additionally, we are
planning to integrate this pose estimation with the additional information from
the walking odometry and information from the IMU. This can be done with the
robot localization package [16] also provided by ROS. This seems to be a very
promising approach in terms of accuracy, as the quality of the walking odometry
is very high due to the new walking engine.

Hamburg Bit-Bots and WF Wolves TDP 2019            7

4.4   Behavior

Our behavior is implemented using the DSD framework (see section 2.3). A
robot can play as goalie, offense or defense player. Based on that general role,
the following behavior differs. While a goalie keeps standing in the goal when
the ball is far away, an offense player walks towards the ball and kicks it in the
direction of the opponents goal. The information on which the behavior is based
is received from the vision and team communication via ROS messages. The
navigation is done with move base [2] which is widely used in mobile robotics.

4.5   Motions

For the full body motion, e.g. kicking and standing up, we are using teach-in
spline animations. While this is similar to the previous years, we improved on
it in two parts. Firstly, we replaced our old cubic spline implementation with a
quintic one. This improves the smoothness and thereby the stability. Secondly,
we improved our teach-in interface and added the possibility to make joints
torque less. Currently we are working on closing the loop on this animations, see
section 2.4.

4.6   Walking

Last year, we participated with an improved version of Rhobans open-loop Quin-
ticWalk engine [14]. With this we were able to walk across the field even with
minor disturbances. While this was already good enough to participate, we are
currently working on improving this, see section 2.5.

5     Conclusion

Since we solved the fundamental problems of the artificial grass and the FIFA
ball last year, we are now focused on improving on stability and performance.
Hamburg Bit-Bots & WF Wolves are looking forward to participating in the
RoboCup 2019 for the Team Competition in Sydney, Australia.

References

 1. mitecom on github, https://github.com/fumanoids/mitecom, accessed:
    02.12.2018
 2. move base - ROS wiki, http://wiki.ros.org/move_base, accessed: 02.12.2018
 3. Rhoban communication board, featuring 9dof IMU and 3 dynamixel
    buses: Rhoban/DXLBoard, https://github.com/Rhoban/DXLBoard, accessed:
    02.12.2018
 4. Robotic foot design involving force sensors, https://github.com/Rhoban/
    ForceFoot, accessed: 02.12.2018

8       Hamburg Bit-Bots and WF Wolves

 5. Bestmann, M., Brandt, H., Engelke, T., Fiedler, N., Gabel, A., Güldenstein, J.,
    Hagge, J., Hartfill, J., Lorenz, T., Heuer, T., Poppinga, M., Salamanca, I.D.R.,
    Speck, D.: Hamburg bit-bots and wf wolves team description for robocup 2018
    humanoid kidsize (2019)
 6. Bestmann, M., Hendrich, N., Wasserfall, F.: Ros for humanoid soccer robots. The
    12th Workshop on Humanoid Soccer Robots at 17th IEEE-RAS International Con-
    ference on Humanoid Robots 2017
 7. Bestmann, M., Reichardt, B., Wasserfall, F.: Hambot: an open source robot for
    robocup soccer. In: Robot Soccer World Cup. pp. 339–346. Springer (2015)
 8. Chitta,     S.,   Marder-Eppstein,      E.,   Meeussen,     W.,     Pradeep,      V.,
    Rodrı́guez Tsouroukdissian, A., Bohren, J., Coleman, D., Magyar, B., Raiola,
    G., Lüdtke, M., Fernández Perdomo, E.: ros control: A generic and simple
    control framework for ros. The Journal of Open Source Software (2017).
    https://doi.org/10.21105/joss.00456, http://www.theoj.org/joss-papers/joss.
    00456/10.21105.joss.00456.pdf
 9. van Dijk, S.G., Scheunemann, M.M.: Deep learning for semantic segmentation on
    minimal hardware. In: Robot World Cup XXII. Springer (2018)
10. Ficht, G., Farazi, H., Brandenburger, A., Rodriguez, D., Pavlichenko, D., Allgeuer,
    P., Hosseini, M., Behnke, S.: Nimbro-op2x: Adult-sized open-source 3d printed
    humanoid robot. In: Robot World Cup XXII. Springer (2018)
11. Fiedler, N., Bestmann, M., Hendrich, N.: Imagetagger: An open source online plat-
    form for collaborative image labeling. In: Robot World Cup XXII. Springer (2018)
12. Gabel, A., Heuer, T., Schiering, I., Gerndt, R.: Jetson, where is the ball? - using
    neural networks for ball detection at robocup 2017. In: Robot World Cup XXII.
    Springer (2018)
13. Gerkey, B.P.: amcl, accessed: 02.12.2018
14. J. Allali, R. Fabre, H.G.L.G.L.H.O.L.S.N.G.P.A.P.Q.R.: Rhoban football club –
    team description paper humanoid kid-size league, robocup 2018 montreal. In:
    Robot World Cup XXII. Springer (2018)
15. Koenig, N.P., Howard, A.: Design and use paradigms for gazebo, an open-source
    multi-robot simulator. In: IROS. vol. 4, pp. 2149–2154. Citeseer (2004)
16. Moore, T.: robot localization. http://docs.ros.org/melodic/api/robot_
    localization/html/index.html, accessed: 02.12.2018
17. Peng, X.B., Abbeel, P., Levine, S., van de Panne, M.: Deepmimic: Example-
    guided deep reinforcement learning of physics-based character skills. arXiv preprint
    arXiv:1804.02717 (2018)
18. Quigley, M., Conley, K., Gerkey, B.P., Faust, J., Foote, T., Leibs, J., Wheeler, R.,
    Ng, A.Y.: Ros: an open-source robot operating system. ICRA Workshop on Open
    Source Software 2009 (2009)
19. Ruppel, P., Hendrich, N., Starke, S., Zhang, J.: Cost functions to specify full-body
    motion and multi-goal manipulation tasks. In: 2018 IEEE International Conference
    on Robotics and Automation (ICRA). pp. 3152–3159. IEEE (2018)
20. Speck, D., Barros, P., Weber, C., Wermter, S.: Ball localization for robocup soccer
    using convolutional neural networks. In: Robot World Cup XX. Springer (2016)
21. Speck, D., Bestmann, M., Barros, P.: Towards real-time ball localization using
    cnns. In: Robot World Cup XXII. Springer (2018)